DeepSeek-R1

DeepSeek-AI · 2025-01-22 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE full-attention text-only open-weights frontier reasoning-specialised

Design overview

DeepSeek-R1 is not an architecture paper: it is a post-training report on a 671B/37B-active MoE model whose base is DeepSeek-V3, already in the atlas. Everything the report discloses about the block fits in Table 4 — 'Architecture: MoE', 671B total, 37B activated — so every deeper spec (attention, experts, context, vocabulary) is deliberately unknown here rather than imported from the V3 record. The substance is the two-stage reasoning recipe: R1-Zero showed pure GRPO with rule-based rewards, no SFT at all, elicits long chain-of-thought and the emergent 'aha moment'; R1 adds a cold-start CoT SFT before RL, then a rejection-sampled SFT round and a second RL stage for helpfulness/harmlessness. The report also established the pattern of distilling reasoning into small dense models, which several later open-weight families in the gallery follow.

What the report claims is novel

Atlas assessment: Architecturally this record adds nothing new to the atlas — the report states the base is DeepSeek-V3 (already a record) and discloses no block details. Its value is as the reasoning-specialised exemplar of the sparse-MoE line: the first major open report showing rule-based RLVR alone (without SFT) can elicit chain-of-thought reasoning, plus the distillation-of-reasoning pattern that several later atlas entries adopt.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyDeepSeek-R1[1]
OrganisationDeepSeek-AI[2]
Release date2025-01-22
Report URLhttps://arxiv.org/abs/2501.12948[3]
Open weightsyes[4]

Scale

Total parameters671B[5]
Active parameters / token37B[6]

Core block

Block typesparse-MoE[7]

Context

Tokenizer

Training

Curriculum / stagingnone disclosed[3]

Post-training

SFTyes[8]
Preference optimisationGRPO[9]
Reasoning trainingTwo-stage: (1) cold-start CoT SFT on V3-Base → GRPO reasoning RL with rule-based accuracy/format rewards (R1-Zero: pure RL from V3-Base, no SFT); (2) rejection-sampled SFT (~800k samples) → RL for all scenarios; <think>/<answer> template; max generation 32,768 tokens; eval temp 0.6, top-p 0.95[10]
Distillationnone

Modality

Typetext-only

Inference efficiency

Quantisation shippednone disclosed[3]
Speculative decodingnone disclosed[3]
Serving optimisationsnone disclosed[3]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
AIME 2024 (Pass@1)79.8Table 3
MATH-500 (Pass@1)97.3Table 3
Codeforces (rating)2029Table 3
MMLU90.8Table 3
MMLU-Pro84.0Table 3
GPQA Diamond71.5Table 3
AlpacaEval2 (LC)87.6Table 3
ArenaHard92.3Table 3
SWE Verified49.2Table 3
LiveCodeBench65.9Table 3

Provenance references

  1. stated · Title
  2. stated · Title page byline
  3. stated · pipeline
  4. stated · Abstract: 'we open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models'
  5. stated · Table 4 (# Total Params 671B)
  6. stated · Table 4 (# Activated Params 37B)
  7. stated · Table 4 (Architecture: MoE; no per-layer pattern disclosed in this report)
  8. stated · §2.3 (cold-start CoT SFT on DeepSeek-V3-Base; rejection-sampled SFT)
  9. stated · §2.2 (reasoning RL uses GRPO with rule-based rewards; second RL stage uses rule + reward-model rewards)
  10. stated · §2.2–2.3